See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/357505893

Model Free Error Compensation for Cable-Driven Robot Based on Deep Learning
with Sim2real Transfer Learning

Chapter · January 2022

DOI: 10.1007/978-3-030-92442-3_24

CITATIONS READS

2 222

5 authors, including:

Aydar Akhmetzyanov Maksim Rassabin

Innopolis University Innopolis University
6 PUBLICATIONS 45 CITATIONS 6 PUBLICATIONS 33 CITATIONS

SEE PROFILE SEE PROFILE

Alexander Maloletov Mikhail Fadeev

Innopolis University Innopolis University
53 PUBLICATIONS 177 CITATIONS 4 PUBLICATIONS 24 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Alexandr Klimchik on 06 November 2022.

The user has requested enhancement of the downloaded file.

 Model free error compensation for cable-driven
robot based on Deep Learning with sim2teal
Transfer Learning

Aydar Akhmetzyanov1 , Maksim Rassabin1 , Alexander Maloletov1 , Mikhail

Fadeev1 , and Alexandr Klimchik1

Innopolis University, Universitetskaya St, 1, Innopolis, Tatarstan, 420500, Russia

[email protected], {m.rassabin, a.maloletov, m.fadeev,
a.klimchik}@innopolis.ru

Abstract. The paper deals with model-free error compensation for cable-
driven parallel robots based on the sim2real deep transfer learning. Par-
ticular attention is paid to simulation-based error estimation for different
payloads attached to the robot end-effector and the use of the Transfer
Learning approach for error compensation. This allows to reduce physi-
cal experiments with a real robot and gather sufficient data set within a
reasonable time, which is required for deep learning. The obtained results
were applied and validated for underactuated 4-dof (degrees of freedom)
cable-driven parallel robot. Model-free Deep learning-based methods for
a considerable training dataset provides better accuracy than simple lin-
ear error compensators using model-based calibration procedure. The
proposed sim2real Transfer Learning method allowed to speed up the
process of robotics system integration and recalibration due to the sig-
nificant sample efficiency improvement.

Keywords: cable-driven robot, sim2real, kinematics compensation, deep

learning, transfer learning, robotics

1 Introduction
Parallel cable-driven robots are a new automation tool for solving a number
of complex technological tasks. They are characterized by high payload, large
workspace for moving mobile platform, and low inertia properties. These primary
advantages allow to use the cable-driven parallel robots for different applications.
Their advantages include larger workspace [32], relatively small robot mass, abil-
ity to handle large payload, ability to operate with high speed and acceleration.
The main disadvantage of parallel cable-controlled robots is associated with the
complexity of physical modeling and, as a result, with the complexity of non-
linear compensation of geometric and non-geometric errors.
From an industrial point of view, the cable-driven robots can be used for:
– moving of heavy loads over long distances [3], [13], [15], [25];
– moving objects at relatively high speeds [16];
2 A. Akhmetzyanov et al.

– creating force feedback systems for virtual reality or for flying in aerotube
[24], [31];
– transportable and deployable systems for fast rescue operations [29].

One of the key moments for the control of such robots is the considering of
the elastostatic deformation and dynamics of the mobile platform, and cabling.
In particular, Zhang and Shang [43] worked for trajectory planning of the three
degrees underactuated cable-driven robot, taking into account the dynamic ef-
fects. In their work, the geometric approach for path planning was proposed,
which can also be applied when the mechanism is beyond the scope of static
equilibrium. Proposed by the authors approach gives an analytical solution that
enables positive and continuous tension in all control cables. Dynamics control
of actuators is also taken into account in [46] for precise control of large spheri-
cal radio telescope. Influence of dynamics on control for cable systems has been
shown in the work of Du et al. [9], where the authors used a dynamic model
with variable cable length for managing a mobile platform. Gosselin et al. [14]
used a dynamic simulation for the trajectory planning of the robot movement
beyond the static stability. The importance of dynamic modeling for the control
of cable robots supported by a large number of scientific papers devoted to this
aspect [34], [45]
Stiffness analysis of cable mechanisms examined in several works [27], [22].
In the work, Carricato and Merle [6] considered the problem of the properties
analysis of the robot in view of the rigidity and geometry. In the work [41] the
variable stiffness actuators were used to provide the desired characteristics of
the cable robot in a workspace. Static and dynamic analysis of cable robots with
the stiffness and weight control were considered in Yuan et al. [42].
One of the biggest challenges associated with cable robots is ensuring desired
properties of the robot in the operating area [36], [37]. Today, cable robots have
the most complex kinematics [28], which makes the task of analyzing the kine-
matic and dynamic properties of the manipulator in the workspace non-trivial
task [12]. The task of synthesis and optimization of cable robots is usually solved
separately for each application and set of requirements for certain properties of
the robot [38].
For more precise control of cable robots, the external measurement systems
can be used. In the paper [5] was presented a control algorithm for the cable
robot, the algorithm was based on machine vision. In the mentioned work, they
considered the kinematic simulation of a robot with four links, static and dy-
namic balancing, online image processing, and a variety of control algorithms.
Authors used an adaptive controller with integrated sliding recursive identifi-
cation to implement more precise control. Alternative control algorithms cable
robot based on vision systems used in the work [8], [7]. Adaptive control and
control based on the PID is also widely used to control the cable robot [4], [17].
In order to improve robot positional accuracy while performing technological
task in an industrial environment different on-line and off-line error compensa-
tion methods are used [40], [18], [20], which are aimed to minimize the influence
of geometric dynamic and elastic factors on the deviation of target position from
 Model free error compensation based on sim2teal learning 3

the real one. In the cable-driven parallel robot, the end-effector position can be
changed only via adjustment cable lengths, which should rely on the correspond-
ing error compensation algorithm. That does not require controller modification
and makes the error compensation algorithm relatively simple for implemen-
tation by means of updating higher-level inputs only. In this case, in order to
achieve the desired positioning accuracy, it is required to give reference trajec-
tory as input that differs from the target one on the errors to be compensated
[21]. The input values are computed either iteratively using the kinematic model
or based on the Jacobian matrix [19].
In general, the error compensation algorithms can be split into two big
groups: based on some sophisticated model [19] or model-free compensation [44].
The first group can be easily adopted within the robot workspace but it is not
able to take into account any factor that is not described by the model. The
second group does not need any preliminary knowledge on the robot, may take
into account all possible factor influences on the robot positioning accuracy, but
frequently requires either a considerable amount of data for training (for machine
learning-based algorithms) or real-time estimation of the end-effector position.
Real-time estimation of the end-effector position is usually not possible on the
industrial floor; that is why this approach is commonly used for some validations
in the lab only. Recent research in kinematic error compensation shows the effi-
ciency of Deep learning and Reinforcement learning methods [35], [2] . The most
straightforward approach to compensate robot positioning errors is based on the
simple linear regression model which provides acceptable results for the majority
of real cases. Better results can be achieved using Deep Learning and Transfer
Learning approaches that are focused on in this work. This work is extension
of previously published work [2] containing additional experimental results and
analysis transfer learning effectiveness for model free error compensation.

2 System Overview

2.1 Cable-driven robot description

The cable-driven robot considered in this work consists of a frame, four winches
with cables, and a mobile platform (Fig. 1). Winch mechanisms are located at
the bottom of the frame. The cables are thrown through the guide rollers in
the upper part of the frame. Mountings of guide rollers can rotate around a
vertical axis, providing an orientation of the cables in the direction of the mobile
platform. The free ends of the cables are attached to a mobile platform on which
specific equipment can be placed.
The 4-cable robot belongs to the class of underactuated system. In this robot,
it is possible to control the position of the mobile platform and single orienta-
tion around virtual axis varying within the work-space. for such a system cable
sagging affects essentially the end-effector positioning accuracy.
Formally, 4 actuators should provide control of 4 degrees of freedom, for ex-
ample, they should allow to control the position of the mobile platform and the
4 A. Akhmetzyanov et al.

Fig. 1. The design scheme of the cable-driven robot

angle of rotation around a certain axis. However, the cables are non-restraining
geometric connections, which are closed only due to external forces acting on
the mobile platform. In other words, cables can only work in tension. There-
fore, using 4 cables, control of 4 generalized coordinates is possible only in such
configurations of the cable system in which one of the cables is an antagonist
for the other three cables, and at the same time its tension is ensured by the
external forces of the system. In the configuration of the cable robot under con-
sideration, the mobile platform is suspended from 4 cables. Therefore, none of
the cables can be an antagonist to the rest. And this means that to control 4
generalized coordinates, at least one of the cables must work in compression,
which is impossible.
The robot control system solves the inverse kinematics problem by deter-
mining the cable lengths according to the given position of the mobile platform,
taking into account the peculiarities of the winding mechanism and the guide
rollers construction [26]. However, this system does not take into account sag-
ging cables and other possible factors affecting the positioning accuracy of the
mobile platform.
The control system provides the ability to enter a compensating factor to
improve the accuracy of robot control [11]. But the problem is the complexity
of a fairly accurate estimate of the value of the compensating factor. Direct
measurement of the positioning error of the mobile platform during the operation
 Model free error compensation based on sim2teal learning 5

of the robot is not always possible. A more accurate dynamic model of the robot,
taking into account many parameters, requires large computational costs during
the operation of the robot and the time required for calibration to determine the
values of these parameters. Accordingly, a high-speed neural network is a good
solution. Due to the Transfer Learning method, the time spent on retraining the
network on a particular robot will be comparable to the time spent on calibrating
the dynamic model.

2.2 Modelling environment: Unity robot simulation

To collect data for transfer learning we created the unity simulation of the cable-
driven parallel robot that integrated the physical engine. The robot structure
screenshot is available in Fig. 2. The robot consists of 4 elastic prismatic joints,
the end-effector, cable end object which connects prismatic joints to the end-
effector. We can control joint lengths and thus move the end-effector inside the
robot workspace. The end-effector mass influences the tension of robot springs
and changes the kinematics of the robot.

Fig. 2. Cable-driven parallel robot structure in the Unity environment

To collect the dataset for the transfer learning experiment we defined a set
of the target position and estimated their end-effector positions in the simulated
environments. The randomly generated joint length allowed to collect the dataset
for different situations. We performed the data collection for 2 different masses
which allowed to estimate the advantage from neural network weight transfer
for the training of the new error compensator.
The resulting dataset consists of the target and the measured positions in the
unity frame for 2 different masses and available under simulated kinematicsMass1.csv
and simulated kinematicsMass2.csv files in the dataset folder [1]. To further ap-
6 A. Akhmetzyanov et al.

ply this dataset to the real robot, the coordinate frame should be converted to
the robot coordinate system.

2.3 Sim2Real Transfer Learning with Deep Learning

Deep learning is related to representation or feature learning term, the process

of finding an appropriate representation of data. In our case, the neural network
can learn inverse kinematics without predefined knowledge of the robot struc-
ture. This method first proved its efficiency on image classification tasks [23]
and later was expanded to other functions including robot kinematics [10]. The
primary drawback of deep learning compensation methods is the requirement
of the extensive training dataset, resulting in increased calibration time. In this
paper, we address the Transfer Learning paradigm to reduce the demand for the
training set with real robots for neural network training.
Transfer learning is the improvement of learning in a new task through the
transfer of knowledge from a related task that has already been learned [39].
Transfer learning also can be characterized as a process of knowledge transfer
from one task to another, which is visualized in Figure 3.

Fig. 3. Traditional ML and transfer learning scheme

Transfer learning is a broad research field. Its application includes the Nat-
ural Language Processing field with word2vec [30], Reinforcement learning with
Curriculum [33], etc. The benefit of transfer learning is visualized in Fig. 4 [39]. It
allows us to improve initial accuracy, improve the slope of the training curve, or
increase the asymptote. In our robotics case, transfer learning can be potentially
used to compensate the impact of different payload application to our robot or
to speed up the process of regular robot re-calibration.
Employing testing the initial idea of applying a transfer learning approach for
cable-driven robot calibration, we modeled the robot in Unity with an integrated
Nvidia PhysX simulation engine. The originality of this work corresponds to the
 Model free error compensation based on sim2teal learning 7

Fig. 4. Learning performance [39]

reduction of training set size required for the neural network through Transfer
Learning application and validation of the proposed approach on 4-dof cable-
driven robot.

3 Implementation and Results

To check the transfer learning viability for robot calibration tasks, we imple-
mented the following experiments. In the first one, we modeled a robot with
Unity and implemented simple inverse kinematics. Using this simulator, we col-
lected the inverse kinematics reference error dataset. In the second one, we
trained the error compensator on it. In the third one, we modified robot pa-
rameters to simulate the uncalibrated robot behavior with additional weight
applied to the end-effector and used transfer learning to speed up the process of
compensatory training. In the fourth experiment, we tried to verify the viability
of transfer learning to train compensators on the data from real robots. Finally,
we applied different weighed payloads on the robot and recalibrated the com-
pensator in a new usage scenario. The results, the dataset, the source code, and
the unity project are available in the GitHub repository [1].

3.1 Unity model compensation

We developed a neural network model that is capable to fit kinematics and

outperform the linear model that has been taken as a baseline. We implemented
our model with Keras framework with a TensorFlow backend. We provide XYZ
coordinates as the input to the network (3 float inputs). Output in our case
is a correction signal for XYZ coordinates (3 float outputs). The best results
were obtained with the neural network model with one hidden layer containing
7 neurons, the hyperbolic tangent activation function, and input normalization.
8 A. Akhmetzyanov et al.

We provided the target position as input and kinematic error as a prediction

target. Thus, our neural network must accept the coordinates of the position
of the mobile platform and provide the predicted value of the positioning error,
which we can use as a compensating factor in the control system of the cable
robot.
We tested sigmoid, linear, tanh, and ReLU activation functions with different
neural network architectures. The Hyperbolic tangent activation function gave
the best results, because of its range symmetry around zero which is true for
linear activation as well.
We provide the visualization of kinematics error data (Fig. 5) as a difference
between reference and measured positions. The color of the dots corresponds
to the kinematic error: the brighter, the worse. There is the co-directed shift in
positions which can be compensated. Errors are not linear relative to the position
in the workspace and thus the non-linear model is required to compensate errors.
MAE for the weight #1 dataset is 25.09 mm, for the weight #2 is 33.93 mm.
Training history is available in Figure 6 [2]. After compensation mean error was
reduced to 0.8 mm.

Fig. 5. Kinematic errors on simulated data

3.2 Unity model transfer learning

Low error for compensator training is possible due to unlimited data availability
from the simulation. It allows us to estimate optimal model architecture. For
 Model free error compensation based on sim2teal learning 9

Fig. 6. Training history on simulated data [2]

further experiments, we will use only a limited subset of available simulation

data. To verify our transfer learning hypothesis, we are comparing sample re-
quirements for model training from scratch and with inherited neural network
weights from the previous payload type. Besides, we will compare both models
with a linear compensator as a baseline.
The results of the experiments are presented in Figure 7 [2]. Here we present
the error distance (training and test set dist) for different training set sizes. The
Hypothesis was confirmed and it seems that weight transfer is more efficient in
terms of sample efficiency. We see that 22 samples are enough to outperform the
linear model. But with transfer learning, 22 are enough to adapt the previous
model to the new task with 2.5 times accuracy increase. The same accuracy can
be achieved with 100 samples for training from scratch. We also increasing the
number of training epochs while increasing the number of training samples to
improve training performance.
In our figures, we use epochs and loss dimensions. In artificial neural network
terms, an epoch stands for one cycle of gradient descent and backpropagation
through the full training dataset. For the loss term, we use mean absolute error
or L1 Loss which value is measured in millimeters, thus, naturally describes the
10 A. Akhmetzyanov et al.

Fig. 7. Train and test errors on Unity data [2]

accuracy of our model. Our loss is the arithmetic mean of absolute differences
between our target reference and predicted values.

3.3 Real robot compensation

The approach developed in this paper is tested on a prototype of the cable-driven

parallel robot presented in Figure 8.
In the experiments, we used a mobile platform, in which all 4 cables are at-
tached to a single point and the position of the mobile platform is determined by
the coordinates of this point. To obtain the real position of the mobile platform,
the FARO Vantage laser tracker is used, which measures the absolute reflector
position with the accuracy of 20 µm + 5 µm/m. To measure the coordinates
of the mobile platform, the reflector was mounted on the platform above the
cables’ attachment point.
Test positions of a mobile platform were measured for 1183 points located in
nodes of a regular grid of 13x13x7 points. Sizes of the investigated working space:
ξ ∈ [-3500 mm, 3500 mm], η ∈ [-1500 mm, 1500 mm], ζ ∈ [0 mm, 1200 mm].
Experimental studies were carried out for 3 different payloads: for the masses
equal to 5, 17 and 33 kg. The difference between the target and obtained positions
of the mobile platform corresponds to the error that should be compensated in
the control loop.
 Model free error compensation based on sim2teal learning 11

Fig. 8. Photo of cable-driven parallel robot prototype used for validation.

We collected more than 2000 samples from real robots for 17 and 33 kg
masses. Visualizations of errors for 33 kg payload available in Figure 9. In the
real robot dataset, the linear shift of error also explicit. For 5 kg payload, we
have a 1.117m mean Z position and 1.085 m for 33 kg, which means that the
bigger mass pulls the end effector down. Our total error for 5kg: 0.198 m and
error for 33kg: 0.186 m. The training process is available in Figure 10 [2]. Test
error is 8.1 mm after training.

3.4 Real robot to real robot with different mass compensation

In the production environment, we need to recalibrate our robot or adapt our

controller for the new payload. This is the common application scenario of our
method. From our experiment, in the case of different mass adaptation for real
robots, transfer learning strongly reasonable. The results are available in Figure
11. This experiment proved that it is efficient to apply transfer learning when
we need to speed up the process of robot calibration or adaptation to a new
end-effector payload.
In Fig. 11 [2] we see that for 20 samples we achieve the same performance for
linear and neural network-based compensators. For the same training set size,
we achieved 30% increase in accuracy using proposed transfer learning method.
To achieve the same accuracy we need 120 samples to train the neural network
from scratch which shows sample efficiency increase by the order of magnitude.
12 A. Akhmetzyanov et al.

Fig. 9. Kinematic error data based on real measurements

3.5 Unity simulation to real robot transfer learning

Our experiments showed that in some cases, it is possible to improve the training
process even when the environment dynamics is different. In our case, we used
 Model free error compensation based on sim2teal learning 13

Fig. 10. Training history with real data [2]

weights from simulated environments to train the real robot error compensator.
As a result, some constraints emerge. For example, the frames should be co-
directed, and scales should be the same. Normalization with zero mean and one
standard deviation helps to achieve this task. Fig. 12 and Fig. 13 show these
results [2].

3.6 Environmentally invariant model

Classical robot modeling involves robot dynamics derivation which incapsulates

environmental hyperparameters such as robot part masses, temperature, payload
mass, etc. Such parameters play a crucial role in the model accuracy and as a
result, allows to adapt the model to altered application with only little effort of
changing those parameters. The same advantage can be implemented for deep
learning-based error compensator as well.
Let’s consider the temperature-invariant model. In industrial robotics, envi-
ronment and temperature, in particular, can play a significant role in the out-
come of robot calibration by changing material properties such as stiffness and
size. In our case of neural-network-based model, such hyperparameter can be
added as an input to the model as an additional dataset feature. In the case of
the temperature-invariant model, the training can be collected with an additional
temperature sensor. As a consequence of such complexity extension, the model
14 A. Akhmetzyanov et al.

Fig. 11. Training and test errors [2]

must be re-engineered with additional neurons or layers to meet the increased

non-linearity requirement.

4 Conclusions

Neural networks is an efficient tool to describe different sources of inaccuracy

influencing on the position of mobile platform of cable-driven robot. For this
type of model-free calibration, a large dataset is required that should contain
all variable parameters with sufficient variation range. Undoubted this creates
a problem when we need to operate with variable mass attached to the robot
end-effector and requires essential variations in robot calibrating experiments
used for machine learning.
Sim2real approach considered in this work used sophisticated but not ex-
act simulation model for fast computations to obtain required data for training
neural network. Subsequently, the neural network is additionally trained using a
small amount of data obtained on a real robot. The Unity framework was used as
a platform for fast robot modeling. The prototype of cable-driven manipulator
with four cables and working area of 7x3x1.2 m was used for experimental study.
The use of simulation and a part of real data, in conjunction with the use of
the transfer learning technique, allows increasing the accuracy of the manipu-
lator operation without increasing data collection. For cable-driven robots, this
 Model free error compensation based on sim2teal learning 15

Fig. 12. Training and test errors for 5 kg [2]

method seems to be especially relevant in connection with the effects of a signif-

icant change in the length of the cables under static load compared to dynamic
load, a thermal cables expansion and other factors due to which it is necessary
to calibrate the cable system every time at the beginning of work. For the pro-
totype of cable-driven robot it was possible to predict a positioning errors with
the preseason 5 mm using 40 different measurements only in a training set.
It is worth mentioning that this work examined the application of the method
using only one type of a manipulator - a cable-driven robot, however, this area
in the field of robotics has prospects and requires further study.

Acknowledgements

Work was supported by the RFBR (Russian Foundation for Basic Research)(Grant
No. 19-08-01234).

References

1. Akhmetzyanov, A.: Github repository. https://github.com/AydarAkhmetzyanov/Deep-

Learning-based-cable-driven-robot-kinematics--error-compensation-with-Transfer-
Learning (2019)
16 A. Akhmetzyanov et al.

Fig. 13. Training and test errors for 33 kg [2]

2. Akhmetzyanov, A., Rassabin, M., Maloletov, A., Fadeev, M., Klimchik, A.:
Deep learning with transfer learning method for error compensation of cable-
driven robot. In: Gusikhin, O., Madani, K., Zaytoon, J. (eds.) Proceed-
ings of the 17th International Conference on Informatics in Control, Au-
tomation and Robotics, ICINCO 2020, Lieusaint, Paris, France, July 7-9,
2020. pp. 553–559. ScitePress (2020). https://doi.org/10.5220/0009905605530559,
https://doi.org/10.5220/0009905605530559
3. Albus, J., Bostelman, R., Dagalakis, N.: The nist spider, a robot crane. Journal
of research of the National Institute of Standards and Technology (97), 373–385
(1992)
4. Babaghasabha, R., Khosravi, M., Taghirad, H.: Adaptive robust control of fully-
constrained cable driven parallel robots. Journal of research of the National Insti-
tute of Standards and Technology (25), 27–36 (2015)
5. Bayani, H., Masouleh, M., Kalhor, A.: An experimental study on the vision-based
control and identification of planar cable-driven parallel robots. Robotics and Au-
tonomous Systems (75, Part B), 187–202 (2016)
6. Carricato, M., Merlet, J.: Geometrico-static analysis of under-constrained cable-
driven parallel robots. Advances in Robot Kinematics: Motion in Man and Machine
p. 309–319 (2010)
7. Dallej, T., Gouttefarde, M., Andreff, N., Dahmouche, R., Martinet, P.: Vision-
based modeling and control of large-dimension cable-driven parallel robots. In:
Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ International Conference
on 2012. pp. 1581–1586 (2012)
 Model free error compensation based on sim2teal learning 17

8. Dallej, T., Gouttefarde, M., Andreff, N., Michelin, M., Martinet, P.: Towards
vision-based control of cable-driven parallel robots. In: Intelligent Robots and Sys-
tems (IROS), 2011 IEEE/RSJ International Conference on 2011. pp. 2855–2860
(2011)
9. Du, J., Bao, H., Cui, C., Yang, D.: Dynamic analysis of cable-driven parallel ma-
nipulators with time-varying cable lengths. Finite Elements in Analysis and Design
(48), 1392–1399 (2012)
10. Duka, A.V.: Neural network based inverse kinematics solution for trajec-
tory tracking of a robotic arm. Procedia Technology 12, 20–27 (12 2014).
https://doi.org/10.1016/j.protcy.2013.12.451
11. Fadeev, M., Maloletov, A.: Control of the four-cable-driven parallel robot with
the help of the inverse kinematic model. In: Proceedings of the 6th International
Young Scientists Conference on Information Technologies, Telecommunications and
Control Systems (ITTCS 2019) (2019)
12. Gao, B., Song, H., Zhao, J., Guo, S., Sun, L., Tang, Y.: Inverse kinematics and
workspace analysis of a cable-driven parallel robot with a spring spine. Mechanism
and Machine Theory (76), 56–69 (2014)
13. Gorman, J., Jablokow, K., Cannon, D.: The cable array robot: theory and experi-
ment. In: Proceedings 2001 ICRA IEEE International Conference on Robotics and
Automation (Cat No01CH37164) 2001. vol. 3, pp. 2804–2810 (2001)
14. Gosselin, C., Ping, R., Foucault, S.: Dynamic trajectory planning of a two-dof
cable-suspended parallel robot. In: Robotics and Automation (ICRA), 2012 IEEE
International Conference on 2012. pp. 1476–1481 (2012)
15. Gouttefarde, M., Collard, J., Riehl, N., Baradat, C.: Simplified static analysis of
large-dimension parallel cable-driven robots. In: 2012 IEEE International Confer-
ence on Robotics and Automation2012. pp. 2299–2305 (2012)
16. Kawamura, S., Choe, W., Tanaka, S., Pandian, S.: Development of an ultrahigh
speed robot falcon using wire drive system. In: Robotics and Automation, 1995
Proceedings, 1995 IEEE International Conference on 1995. vol. 1, pp. 215–220
(1995)
17. Khosravi, M., Taghirad, H.: Robust pid control of fully-constrained cable driven
parallel robots. Mechatronics (24), 87–97 (2014)
18. Klimchik, A., Ambiehl, A., Garnier, S., Furet, B., Pashkevich, A.: Ex-
perimental study of robotic-based machining. IFAC-PapersOnLine 49(12),
174 – 179 (2016). https://doi.org/https://doi.org/10.1016/j.ifacol.2016.07.591,
http://www.sciencedirect.com/science/article/pii/S2405896316308515, 8th IFAC
Conference on Manufacturing Modelling, Management and Control MIM 2016
19. Klimchik, A., Bondarenko, D., Pashkevich, A., Briot, S., Furet, B.: Compli-
ance error compensation in robotic-based milling. CoRR abs/1409.6231 (2014),
http://arxiv.org/abs/1409.6231
20. Klimchik, A., Pashkevich, A.: Robotic manipulators with double en-
coders: accuracy improvement based on advanced stiffness model-
ing and intelligent control. IFAC-PapersOnLine 51(11), 740 – 745
(2018). https://doi.org/https://doi.org/10.1016/j.ifacol.2018.08.407,
http://www.sciencedirect.com/science/article/pii/S2405896318315337, 16th
IFAC Symposium on Information Control Problems in Manufacturing INCOM
2018
21. Klimchik, A., Pashkevich, A., Chablat, D., Hovland, G.: Compliance er-
ror compensation technique for parallel robots composed of non-perfect
serial chains. Robotics and Computer-Integrated Manufacturing 29(2), 385
18 A. Akhmetzyanov et al.

– 393 (2013). https://doi.org/https://doi.org/10.1016/j.rcim.2012.09.008,

http://www.sciencedirect.com/science/article/pii/S0736584512001111
22. Kozak, K., Qian, Z., Jinsong, W.: Static analysis of cable-driven manipulators with
non-negligible cable mass. In: IEEE Transactions on Robotics. vol. 22, pp. 425–433
(2006)
23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep
convolutional neural networks. Commun. ACM 60(6), 84–90 (May 2017).
https://doi.org/10.1145/3065386, https://doi.org/10.1145/3065386
24. Lafourcade, P.: Étude des manipulateurs parallèles à câbles: conception d’une sus-
pension active pour soufflerie (2004)
25. Lamaury, J., Gouttefarde, M.: Control of a large redundantly actuated cable-
suspended parallel robot. In: 2013 IEEE International Conference on Robotics
and Automation 2013. pp. 4659–4664 (2013)
26. Maloletov, A., Fadeev, M., Klimchik, A.: Error analysis in solving the in-
verse problem of the cable-driven parallel underactuated robot kinematics and
methods for their elimination. IFAC-PapersOnLine 52, 1156–1161 (01 2019).
https://doi.org/10.1016/j.ifacol.2019.11.352
27. Merlet, J.: Analysis of wire elasticity for wire-driven parallel robots. In: Proceedings
of EUCOMES 08. pp. 471–478. Springer Netherlands (2009)
28. Merlet, J.: Kinematic analysis of the 4-3-1 and 3-2-1 wire-driven parallel crane. In:
Robotics and Automation (ICRA), 2013 IEEE International Conference on 2013.
pp. 4635–4640 (2013)
29. Merlet, J., Daney, D.: A portable, modular parallel wire crane for rescue operations.
In: Robotics and Automation (ICRA), 2010 IEEE International Conference on
2010. pp. 2834–2839 (2010)
30. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word repre-
sentations in vector space (2013)
31. Morizono, T., Kurahashi, K., Kawamura, S.: Analysis and control of a force display
system driven by parallel wire mechanism. Robotica (16), 551–563 (1998)
32. Morris, M., Shoham, M.: Applications and theoretical issues of cable-driven robots
(05 2009)
33. Narvekar, S., Stone, P.: Learning curriculum policies for reinforcement learning
(2018)
34. Ou, M., Xiumin, D.: Dynamics analysis of a cable-driven parallel manipulator for
hardware-in-the-loop dynamic simulation. In: Advanced Intelligent Mechatronics
Proceedings, 2005 IEEE/ASME International Conference on 2005. pp. 837–842
(2005)
35. Pane, Y., Nageshrao, S., Kober, J., Babuska, R.: Reinforcement learn-
ing based compensation methods for robot manipulators. Engineer-
ing Applications of Artificial Intelligence 78, 236–247 (02 2019).
https://doi.org/10.1016/j.engappai.2018.11.006
36. Pham, C., Yeo, S., Yang, G., Chen, I.: Workspace analysis of fully restrained cable-
driven manipulators. Robotics and Autonomous Systems (57), 901–912 (2009)
37. Pusey, J., Fattah, A., Agrawal, S., Messina, E.: Design and workspace analysis of a
6–6 cable-suspended parallel robot. Mechanism and Machine Theory (39), 761–778
(2004)
38. Seriani, S., Seriani, M., Gallina, P.: Workspace optimization for a planar cable-
suspended direct-driven robot. Robotics and Computer-Integrated Manufacturing
(34), 1–7 (2015)
39. Torrey, L., Shavlik, J., Walker, T., Maclin, R.: Transfer Learning via Advice Taking,
vol. 262, pp. 147–170 (12 2009). https://doi.org/10.1007/978-3-642-05177-7 7
 Model free error compensation based on sim2teal learning 19

40. Wu, Y., Klimchik, A., Caro, S., Furet, B., Pashkevich, A.: Geometric cal-
ibration of industrial robots using enhanced partial pose measurements and
design of experiments. Robotics and Computer-Integrated Manufacturing 35,
151 – 168 (2015). https://doi.org/https://doi.org/10.1016/j.rcim.2015.03.007,
http://www.sciencedirect.com/science/article/pii/S0736584515000411
41. Yeo, S., Yang, G., Lim, W.: Design and analysis of cable-driven manipulators with
variable stiffness. Mechanism and Machine Theory (69), 230–244 (2013)
42. Yuan, H., Courteille, E., Deblaise, D.: Static and dynamic stiffness analyses of
cable-driven parallel robots with non-negligible cable mass and elasticity. Mecha-
nism and Machine Theory (85), 64–81 (2015)
43. Zhang, N., Shang, W.: Dynamic trajectory planning of a 3-dof under-constrained
cable-driven parallel robot. Mechanism and Machine Theory (98), 21–35 (2016)
44. Zhao, G., Zhang, P., Ma, G., Xiao, W.: System identification of the
nonlinear residual errors of an industrial robot using massive mea-
surements. Robotics and Computer-Integrated Manufacturing 59, 104
– 114 (2019). https://doi.org/https://doi.org/10.1016/j.rcim.2019.03.007,
http://www.sciencedirect.com/science/article/pii/S0736584518306008
45. Zi, B., Duan, B., Du, J., Bao, H.: Dynamic modeling and active control of a cable-
suspended parallel robot. Mechatronics (18), 1–12 (2008)
46. Zi, B., Zhu, Z., Du, J.: Analysis and control of the cable-supporting system includ-
ing actuator dynamics. Control Engineering Practice (19), 491–501 (2011)

View publication stats